A scaling study of LaBraM, one of the first EEG foundation models, by Shanghai Jiao Tong University (Bao-Liang Lu's group) and Microsoft Research Asia, with Fudan. It keeps LaBraM's encoder and pretraining objective and scales both sides: ten encoders from 5.8M to 1.97B parameters, pretrained on three EEG collections of 25,067 to 79,910 hours. Across 23 downstream tasks, the best fine-tuned encoder (Atom-L) reaches 59.05% mean balanced accuracy, 3.61 points above the strongest of 16 baseline EEG foundation models. A different encoder (Quark-G) leads linear probing at 54.33%, by 1.24 points.

The more useful findings concern how scale behaves. Larger encoders help full fine-tuning more consistently than linear probing, and a better probe does not predict a better fine-tune. Data-fraction runs suggest saturation that depends on capacity, and downstream accuracy can fall while pretraining loss is still dropping. A shared Qwen3 readout that classifies all tasks jointly reaches 50.50% with Qwen3-8B, and a larger language model gives no consistent gain. Code, checkpoints and a project page are released.

Paper

sciencefoundation-modelscalingeeg